A New View on Planning in Online Reinforcement Learning

TL;DR

This paper introduces goal-space planning (GSP), a novel approach in online reinforcement learning that constrains background planning to subgoals, improving efficiency and learning speed by avoiding inaccurate model issues.

Contribution

The paper proposes GSP, a new method that uses subgoal-conditioned models for more efficient, accurate, and abstracted planning in reinforcement learning, bypassing the need for transition dynamics.

Findings

GSP accelerates learning across various domains.

It effectively propagates value in an abstract goal space.

GSP outperforms traditional model-based methods in efficiency.

Abstract

This paper investigates a new approach to model-based reinforcement learning using background planning: mixing (approximate) dynamic programming updates and model-free updates, similar to the Dyna architecture. Background planning with learned models is often worse than model-free alternatives, such as Double DQN, even though the former uses significantly more memory and computation. The fundamental problem is that learned models can be inaccurate and often generate invalid states, especially when iterated many steps. In this paper, we avoid this limitation by constraining background planning to a set of (abstract) subgoals and learning only local, subgoal-conditioned models. This goal-space planning (GSP) approach is more computationally efficient, naturally incorporates temporal abstraction for faster long-horizon planning and avoids learning the transition dynamics entirely. We show that our GSP algorithm can propagate value from an abstract space in a manner that helps a variety of base learners learn significantly faster in different domains.